I did solve my immediate problem at hand, but now I need to understand why it is solved. ;-)
So here I have couple more questions.
Suppose I have a class that is being exported from the DLL. Now this DLL should be loaded into the memory every time I call:
MyExportedClass *pb = new MyExportedClass;
and it should stay in memory and becomes unloaded only when I call:
delete pb;
Is this correct?
If I understood this correctly and the answer to the previous question is yes, then what should happen in the following scenario:
I have an interface which exported from the dll (dll1) and I have its implementation which is exported from another dll (dll2). So every time I execute:
MyInterface *pInterface = new MyImplementation;
both those dlls should be loaded in memory and they should stay in memory until I call:
delete pInterface;
Is this correct?
Now, if the answer to this question is yes - do I have a control/saying, which library will be unloaded first and which one will be second? Or the unload will always happen right after calling destructor of the appropriate class?
Now, is there a tool which checks if the library becomes unloaded and at which point? I can probably just use the fake DllMain() and check its process_detach case, but my impression always was: use DllMain when the library exports function and don't use DllMain when the library exports classes. I used this approach since MSVC 5/6 (following one of the books about C++).
Was I wrong and I can still use DLLMain in both cases?
Thank you.
DLL loading can be automatic (if listed in the import table) or manual (using LoadLibrary). Some managed frameworks, like .NET, will silently call LoadLibrary for you when they see a DLL import declaration in their metadata, but C++ is not one of these. The closest thing C++ has is delay-loading, where the function that calls LoadLibrary is (by default, you can substitute your own) provided by the compiler.
On the other hand, DLL unloading is always manual. Deleting an object never implicitly unloads a DLL. You have to call FreeLibrary (or FreeLibraryAndExitThread). And you had better not call FreeLibrary while objects defined in that library are still in use.
Now, the COM system in Windows is a bit more complicated, because it manages DLL lifetime for you. But DLL lifetime is still not controlled by object deletion, rather by calling a DllCanUnloadNow function in the DLL. Usually you need to maintain a count of active objects in order to write that function correctly. But it still requires your manual implementation, and you can simply always return false to never unload (this is a bit of a pain during development because you have to close the entire application to try a new plugin version, etc... but actually freeing every usage of a DLL and successfully unloading it is rare anyway)
And certainly there is nothing that will automatically unload a DLL listed in the import table, those get loaded during process startup and never unloaded, no matter how many object instances are created or destroyed.
A follow-up question to Memory leaks when calling ITK from Visual Studio DLL
I refined the problem to the simplest example.
struct A
{
public:
A()
{
mp_data = new int(0x42);
}
~A()
{
delete mp_data;
}
int* mp_data;
};
A a;
When such a global class is defined in a DLL, Visual Studio debug CRT reports that mp_data is leaked on application shutdown. Does anybody know a workaround except disabling leak reporting?
If you are calling _CrtDumpMemoryLeaks() at the end of the main function the behaviour is expected, since mp_data will be deleted after _CrtDumpMemoryLeaks() was called.
You would need to call _CrtDumpMemoryLeaks() after the last destructor of you static objects has been called (or rather in the last destructor after the memory has been freed) if you don't want to see these leaks (quite a difficult task, I wouldn't try it).
The cleaner approach is to allocate all your static objects on the heap instead (at the beginning of main), and de-allocate them at the end of main, and then you can call _CrtDumpMemoryLeaks() and won't see any memory leaks.
FYI static objects with constructors and destructors are considered bad anyways, because the order in which they are constructed/destructed is not deterministic, and because of that static objects often introduce bugs which can't be debugged easily.
Edit regarding Andrey's comment:
You could try to deactivate the automatic call to _CrtDumpMemoryLeaks by calling _CrtSetDbgFlag to unset the _CRTDBG_LEAK_CHECK_DF flag. If that works, you can add a static object which calls _CrtDumpMemoryLeaks() in its destructor. To make sure that this object is destructed last, you can use the #pragma init_seg(compiler) directive.
No clue if this will work... Other than that, all other solutions will most likely require you to modify the ITK library (which should be possible, it's an open source library after all?!).
Any of following solves the problem.
(1) Create a fake dependency of the DLL on MFC, or
(2) Use the solution suggested by smerlin: add this code next to DllMain
struct _DEBUG_STATE
{
_DEBUG_STATE() {}
~_DEBUG_STATE() { _CrtDumpMemoryLeaks(); }
};
#pragma init_seg(compiler)
_DEBUG_STATE ds;
I hit the same symptom in the course of migrating an internal library from static linking to load-time dynamic linking, and it turned out that the problem in my case was that the DLL project and the EXE project were linked to different versions of VC++'s runtime/MFC libraries (one was MBCS and one was Unicode).
In my case, the application and library were both using MFC, and the _AFX_DEBUG_STATE destructor which activates the CRT memory leak dump was being called twice, for two separate objects -- since the DLL and EXE linked to different runtime DLLs, static state in the runtime was effectively duplicated. One of the DLLs would unload and dump leaks too early and show a bunch of false leaks. Switching both projects to use the same character set resolved the separate runtime linkage and also resolved the false leak reports.
In my case, linkage to the two separate runtimes was unintentional and may have caused other problems anyway. This obviously wouldn't be the case when consuming third party libraries with a well-defined ABI where you have no control over what CRT the library is linked to.
Not sure if this would've be applicable in your case but I wanted to post in case it's helpful to others.
In MFC applications you can disable automatic memory leak dump by calling:
AfxEnableMemoryLeakDump(FALSE);
This is supported since Visual Studio 2010. For the documentation, see here.
I had the same problems in Visual Studio 2015. I tried all the solutions. The first solution with fake-MFC-dependency only worked if you choose compiler option /MT in your Dll. So your Dll and main application won't share the same heap. But often /MD is needed, e.g. if you want to share STL container or string objects between the Dll and your main application (Dll-boundary). If /MD is used, app and Dll use the same heap. So the first solution with fake-MFC-dependency didn't work for me.
I don't like the second solution by disabling memory-leak-detection in the main application. When you don't need the Dll with this call in destructor anymore, you have to remember to reactivate the memory-leak-detection in your application.
I found another solution so I didn't have false-memory-leaks anymore.
You only have to use the /delayload linker option for your Dll! That's all :-). It worked for me also with compiler option /MD.
Here you can read something about Dll-boundary (why to use /MD?).
And here you can read something about CRT compiler options in general.
I have a single class which does all the required initialization.
currently i have declared a global object of this class type, which is being instantiated on library load.
I've seen other ways, like delaring
BOOL APIENTRY DllMain
entry point for the shared library, and does the actual initialization on process attach.
does this differ from letting the implicit global initialization to its job? which way is better?
This is what happens during C++ DLL startup:
System calls DLL's entry point, generated by you compiler
Entry point calls DllMainCRTStartup (name may differ), which initializes C/C++ runtimes and instantiates all global objects.
DllMainCRTStartup then calls user-defined DllMain.
I personally prefer DllMain, because this way I can explicitly control order of initialization. When you use global objects in different compilation units, they will be initialized in random order which may bring some unexpected surprises 10 minutes before the deadline.
DllMain also let's you do per-thread initialization, which you can not achieve with global objects. However, it is not portable to other platforms.
P.S. You do NOT need mutex in DllMain, as all calls to it are already serialized under process-global critical section. I.e. it is guaranteed two threads will not enter it at the same time for any purpose. This is also the reason why you should not communicate with other threads, load other libraries etc. from this function; see MSDN article for explanation.
A couple of things that should never be done from DllMain:
Call LoadLibrary or LoadLibraryEx (either directly or indirectly). This can cause a deadlock or a crash.
Synchronize with other threads. This can cause a deadlock.
Acquire a synchronization object that is owned by code that is waiting to acquire the loader lock. This can cause a deadlock.
Initialize COM threads by using CoInitializeEx. Under certain conditions, this function can call LoadLibraryEx.
Call the registry functions. These functions are implemented in Advapi32.dll. If Advapi32.dll is not initialized before your DLL, the DLL can access uninitialized memory and cause the process to crash.
Call CreateProces. Creating a process can load another DLL.
Call ExitThread. Exiting a thread during DLL detach can cause the loader lock to be acquired again, causing a deadlock or a crash.
Call CreateThread. Creating a thread can work if you do not synchronize with other threads, but it is risky.
Create a named pipe or other named object (Windows 2000 only). In Windows 2000, named objects are provided by the Terminal Services DLL. If this DLL is not initialized, call to the DLL can cause the process to crash.
Use the memory management function from the dynamic C Run-Time (CRT). If the CRT DLL is not initialized, calls to these functions can cause the process to crash.
Call functions in User32.dll or Gdi32.dll. Some functions load another DLL, which may not be initialized.
Use managed code.
You need a static boolean initialization variable and a mutex. Statically initialize "initialized" to 0. In your DllMain(), make a call to CreateMutex(). Use bInitialOwner=0 and a unique name for lpName that's unique to your application. Then use WaitForSingleObject() to wait for the mutex. Check if initialized is non-zero. If not, do your initialization, and then set initialized to 1. If initialized is non-zero, do nothing. Finally, release the mutex using ReleaseMutex() and close it using CloseHandle().
Here's some pseudo-code, with error and exception handling omitted:
initialized = 0;
DllMain()
{
mutex = CreateMutex(..., 0, "some-unique-name");
result = WaitForSingleObject(handle, ...);
if (result == WAIT_OBJECT_0) {
if (!initialized) {
// initialization goes here
initialized = 1;
}
}
ReleaseMutex(mutex);
CloseHandle(mutex);
}
hi i would recommend u to prefer a signleton class where u can only create a single object of a class and use it. Sigleton class can be created with a private constructor. Now suppose ur class A is a singleton class its object can be used in a constructor of each Class which u want to initialize. Please give us some sample code so other may help u better.
What's the best way for loading a dll from a dll ?
My problem is I can't load a dll on process_attach, and I cannot load the dll from the main program, because I don't control the main program source. And therefore I cannot call a non-dllmain function, too.
After all the debate that went on in the comments, I think that it's better to summarize my positions in a "real" answer.
First of all, it's still not clear why you need to load a dll in DllMain with LoadLibrary. This is definitely a bad idea, since your DllMain is running inside another call to LoadLibrary, which holds the loader lock, as explained by the documentation of DllMain:
During initial process startup or after a call to LoadLibrary, the system scans the list of loaded DLLs for the process. For each DLL that has not already been called with the DLL_PROCESS_ATTACH value, the system calls the DLL's entry-point function. This call is made in the context of the thread that caused the process address space to change, such as the primary thread of the process or the thread that called LoadLibrary. Access to the entry point is serialized by the system on a process-wide basis. Threads in DllMain hold the loader lock so no additional DLLs can be dynamically loaded or initialized.
The entry-point function should perform only simple initialization or termination tasks. It must not call the LoadLibrary or LoadLibraryEx function (or a function that calls these functions), because this may create dependency loops in the DLL load order. This can result in a DLL being used before the system has executed its initialization code. Similarly, the entry-point function must not call the FreeLibrary function (or a function that calls FreeLibrary) during process termination, because this can result in a DLL being used after the system has executed its termination code.
(emphasis added)
So, this on why it is forbidden; for a clear, more in-depth explanation, see this and this, for some other examples about what can happen if you don't stick to these rules in DllMain see also some posts in Raymond Chen's blog.
Now, on Rakis answer.
As I already repeated several times, what you think that is DllMain, isn't the real DllMain of the dll; instead, it's just a function that is called by the real entrypoint of the dll. This one, in turn, is automatically took by the CRT to perform its additional initialization/cleanup tasks, among which there is the construction of global objects and of the static fields of the classes (actually all these from the compiler's perspective are almost the same thing). After (or before, for the cleanup) it completes such tasks, it calls your DllMain.
It goes somehow like this (obviously I didn't write all the error checking logic, it's just to show how it works):
/* This is actually the function that the linker marks as entrypoint for the dll */
BOOL WINAPI CRTDllMain(
__in HINSTANCE hinstDLL,
__in DWORD fdwReason,
__in LPVOID lpvReserved
)
{
BOOL ret=FALSE;
switch(fdwReason)
{
case DLL_PROCESS_ATTACH:
/* Init the global CRT structures */
init_CRT();
/* Construct global objects and static fields */
construct_globals();
/* Call user-supplied DllMain and get from it the return code */
ret = DllMain(hinstDLL, fdwReason, lpvReserved);
break;
case DLL_PROCESS_DETACH:
/* Call user-supplied DllMain and get from it the return code */
ret = DllMain(hinstDLL, fdwReason, lpvReserved);
/* Destruct global objects and static fields */
destruct_globals();
/* Destruct the global CRT structures */
cleanup_CRT();
break;
case DLL_THREAD_ATTACH:
/* Init the CRT thread-local structures */
init_TLS_CRT();
/* The same as before, but for thread-local objects */
construct_TLS_globals();
/* Call user-supplied DllMain and get from it the return code */
ret = DllMain(hinstDLL, fdwReason, lpvReserved);
break;
case DLL_THREAD_DETACH:
/* Call user-supplied DllMain and get from it the return code */
ret = DllMain(hinstDLL, fdwReason, lpvReserved);
/* Destruct thread-local objects and static fields */
destruct_TLS_globals();
/* Destruct the thread-local CRT structures */
cleanup_TLS_CRT();
break;
default:
/* ?!? */
/* Call user-supplied DllMain and get from it the return code */
ret = DllMain(hinstDLL, fdwReason, lpvReserved);
}
return ret;
}
There isn't anything special about this: it also happens with normal executables, with your main being called by the real entrypoint, which is reserved by the CRT for the exact same purposes.
Now, from this it will be clear why the Rakis' solution isn't going to work: the constructors for global objects are called by the real DllMain (i.e. the actual entrypoint of the dll, which is the one about the MSDN page on DllMain talks about), so calling LoadLibrary from there has exactly the same effect as calling it from your fake-DllMain.
Thus, following that advice you'll obtain the same negative effects of calling directly LoadLibrary in the DllMain, and you'll also hide the problem in a seemingly-unrelated position, which will make the next maintainer work hard to find where this bug is located.
As for delayload: it may be an idea, but you must be really careful not to call any function of the referenced dll in your DllMain: in fact, if you did that you would trigger a hidden call to LoadLibrary, which would have the same negative effects of calling it directly.
Anyhow, in my opinion, if you need to refer to some functions in a dll the best option is to link statically against its import library, so the loader will automatically load it without giving you any problem, and it will resolve automatically any strange dependency chain that may arise.
Even in this case you mustn't call any function of this dll in DllMain, since it's not guaranteed that it's already been loaded; actually, in DllMain you can rely only on kernel32 being loaded, and maybe on dlls you're absolutely sure that your caller has already loaded before the LoadLibrary that is loading your dll was issued (but still you shouldn't rely on this, because your dll may also be loaded by applications that don't match these assumptions, and just want to, e.g., load a resource of your dll without calling your code).
As pointed out by the article I linked before,
The thing is, as far as your binary is concerned, DllMain gets called at a truly unique moment. By that time OS loader has found, mapped and bound the file from disk, but - depending on the circumstances - in some sense your binary may not have been "fully born". Things can be tricky.
In a nutshell, when DllMain is called, OS loader is in a rather fragile state. First off, it has applied a lock on its structures to prevent internal corruption while inside that call, and secondly, some of your dependencies may not be in a fully loaded state. Before a binary gets loaded, OS Loader looks at its static dependencies. If those require additional dependencies, it looks at them as well. As a result of this analysis, it comes up with a sequence in which DllMains of those binaries need to be called. It's pretty smart about things and in most cases you can even get away with not following most of the rules described in MSDN - but not always.
The thing is, the loading order is unknown to you, but more importantly, it's built based on the static import information. If some dynamic loading occurs in your DllMain during DLL_PROCESS_ATTACH and you're making an outbound call, all bets are off. There is no guarantee that DllMain of that binary will be called and therefore if you then attempt to GetProcAddress into a function inside that binary, results are completely unpredictable as global variables may not have been initialized. Most likely you will get an AV.
(again, emphasis added)
By the way, on the Linux vs Windows question: I'm not a Linux system programming expert, but I don't think that things are so different there in this respect.
There are still some equivalents of DllMain (the _init and _fini functions), which are - what a coincidence! - automatically took by the CRT, which in turn, from _init, calls all the constructors for the global objects and the functions marked with __attribute__ constructor (which are somehow the equivalent of the "fake" DllMain provided to the programmer in Win32). A similar process goes on with destructors in _fini.
Since _init too is called while the dll loading is still taking place (dlopen didn't return yet), I think that you're subject to similar limitations in what you can do in there. Still, in my opinion on Linux the problem is felt less, because (1) you have to explicitly opt-in for a DllMain-like function, so you aren't immediately tempted to abuse of it and (2), Linux applications, as far as I saw, tend to use less dynamic loading of dlls.
In a nutshell
No "correct" method will allow you to reference to any dll other than kernel32.dll in DllMain.
Thus, don't do anything important from DllMain, neither directly (i.e. in "your" DllMain called by the CRT) neither indirectly (in global class/static fields constructors), especially don't load other dlls, again, neither directly (via LoadLibrary) neither indirectly (with calls to functions in delay-loaded dlls, which trigger a LoadLibrary call).
The right way to have another dll loaded as a dependency is to - doh! - mark it as a static dependency. Just link against its static import library and reference at least one of its functions: the linker will add it to the dependency table of the executable image, and the loader will load it automatically (initializing it before or after the call to your DllMain, you don't need to know about it because you mustn't call it from DllMain).
If this isn't viable for some reason, there's still the delayload options (with the limits I said before).
If you still, for some unknown reason, have the inexplicable need to call LoadLibrary in DllMain, well, go ahead, shoot in your foot, it's in your faculties. But don't tell me I didn't warn you.
I was forgetting: another fundamental source of information on the topic is the [Best Practices for Creating DLLs][6] document from Microsoft, which actually talks almost only about the loader, DllMain, the loader lock and their interactions; have a look at it for additional information on the topic.
Addendum
No, not really an answer to my question. All it says is: "It's not possible with dynamic linking, you must link statically", and "you musn't call from dllmain".
Which *is* an answer to your question: under the conditions you imposed, you can't do what you want. In a nutshell of a nutshell, from DllMain you can't call *anything other than kernel32 functions*. Period.
Although in detail, but I'm not really interested in why it doesn't work,
You should, instead, because understanding why the rules are made in that way makes you avoid big mistakes.
fact is, the loader is not resolving dependenies correctly and the loading process is improperly threaded from Microsoft's part.
No, my dear, the loader does its job correctly, because *after* LoadLibrary has returned, all the dependencies are loaded and everything is ready to be used. The loader tries to call the DllMain in dependency order (to avoid problems with broken dlls which rely on other dlls in DllMain), but there are cases in which this is simply impossible.
For example, there may be two dlls (say, A.dll and B.dll) that depend on each other: now, whose DllMain is to call first? If the loader initialized A.dll first, and this, in its DllMain, called a function in B.dll, anything could happen, since B.dll isn't initialized yet (its DllMain hasn't been called yet). The same applies if we reverse the situation.
There may be other cases in which similar problems may arise, so the simple rule is: don't call any external functions in DllMain, DllMain is just for initializing the internal state of your dll.
The problem is there is no other way then doing it on dll_attach, and all the nice talk about not doing anything there is superfluous, because there is no alternative, at least not in my case.
This discussion is going on like this: you say "I want to solve an equation like x^2+1=0 in the real domain". Everybody says you that it's not possible; you say that it's not an answer, and blame the math.
Someone tells you: hey, you can, here's a trick, the solution is just +/-sqrt(-1); everybody downvotes this answer (because it's wrong for your question, we're going outside the real domain), and you blame who downvotes. I explain you why that solution is not correct according to your question and why this problem can't be solved in the real domain. You say that you don't care about why it can't be done, that you can only do that in the real domain and again blame math.
Now, since, as explained and restated a million times, under your conditions your answer has no solution, can you explain us why on earth do you "have" to do such an idiotic thing as loading a dll in DllMain? Often "impossible" problems arise because we've chosen a strange route to solve another problem, which brings us to deadlock. If you explained the bigger picture, we could suggest a better solution to it which does not involve loading dlls in DllMain.
PS: If I statically link DLL2 (ole32.dll, Vista x64) against DLL1 (mydll), which version of the dll will the linker require on older operating systems?
The one that is present (obviously I'm assuming you're compiling for 32 bit); if an exported function needed by your application isn't present in the found dll, your dll is simply not loaded (LoadLibrary fails).
Addendum (2)
Positive on injection, with CreateRemoteThread if you wanna know. Only on Linux and Mac the dll/shared library is loaded by the loader.
Adding the dll as a static dependency (what has been suggested since the beginning) makes it to be loaded by the loader exactly as Linux/Mac do, but the problem is still there, since, as I explained, in DllMain you still cannot rely on anything other than kernel32.dll (even if the loader in general intelligent enough to init first the dependencies).
Still, the problem can be solved. Create the thread (that actually calls LoadLibrary to load your dll) with CreateRemoteThread; in DllMain use some IPC method (for example named shared memory, whose handle will be saved somewhere to be closed in the init function) to pass to the injector program the address of the "real" init function that your dll will provide. DllMain then will exit without doing anything else. The injector application, instead, will wait for the end of the remote thread with WaitForSingleObject using the handle provided by CreateRemoteThread. Then, after the remote thread will be ended (thus LoadLibrary will be completed, and all the dependencies will be initialized), the injector will read from the named shared memory created by DllMain the address of the init function in the remote process, and start it with CreateRemoteThread.
Problem: on Windows 2000 using named objects from DllMain is prohibited because
In Windows 2000, named objects are provided by the Terminal Services DLL. If this DLL is not initialized, calls to the DLL can cause the process to crash.
So, this address may have to be passed in another manner. A quite clean solution would be to create a shared data segment in the dll, load it both in the injector application and in the target one and have it put in such data segment the required address. The dll would obviously have to be loaded first in the injector and then in the target, because otherwise the "correct" address would be overwritten.
Another really interesting method that can be done is to write in the other process memory a little function (directly in assembly) that calls LoadLibrary and returns the address of our init function; since we wrote it there, we can also call it with CreateRemoteThread because we know where it is.
In my opinion, this is the best approach, and is also the simplest, since the code is already there, written in this nice article. Have a look at it, it is quite interesting and it probably will do the trick for your problem.
The most robust way is to link the first DLL against the import lib of the second. This way, the actual loading of the second DLL will be done by Windows itself. Sounds very trivial, but not everyone knows that DLLs can link against other DLLs. Windows can even deal with cyclic dependencies. If A.DLL loads B.DLL which needs A.DLL, the imports in B.DLL are resolved without loading A.DLL again.
I suggest you to use delay-loading mechanism. The DLL will be loaded at the fisrt time you call imported function. Moreover you can modify load function and error handling. See Linker Support for Delay-Loaded DLLs for more info.
One possible answer is through the use of LoadLibrary and GetProcAddress to access pointers to functions found/located inside the loaded dll - but your intentions/needs aren't clear enough to determine if this is a suitable answer.
Let's say I write a DLL in C++, and declare a global object of a class with a non-trivial destructor. Will the destructor be called when the DLL is unloaded?
In a Windows C++ DLL, all global objects (including static members of classes) will be constructed just before the calling of the DllMain with DLL_PROCESS_ATTACH, and they will be destroyed just after the call of the DllMain with DLL_PROCESS_DETACH.
Now, you must consider three problems:
0 - Of course, global non-const objects are evil (but you already know that, so I'll avoid mentionning multithreading, locks, god-objects, etc.)
1 - The order of construction of objects or different compilation units (i.e. CPP files) is not guaranteed, so you can't hope the object A will be constructed before B if the two objects are instanciated in two different CPPs. This is important if B depends on A. The solution is to move all global objects in the same CPP file, as inside the same compilation unit, the order of instanciation of the objects will be the order of construction (and the inverse of the order of destruction)
2 - There are things that are forbidden to do in the DllMain. Those things are probably forbidden, too, in the constructors. So avoid locking something. See Raymond Chen's excellent blog on the subject:
Some reasons not to do anything scary in your DllMain
Another reason not to do anything scary in your DllMain: Inadvertent deadlock
Some reasons not to do anything scary in your DllMain, part 3
In this case, lazy initialization could be interesting: The classes remain in an "un-initialized" state (internal pointers are NULL, booleans are false, whatever) until you call one of their methods, at which point they'll initialize themselves. If you use those objects inside the main (or one of the main's descendant functions), you'll be ok because they will be called after execution of DllMain.
3 - Of course, if some global objects in DLL A depend on global objects in DLL B, you should be very very careful about DLL loading order, and thus dependancies. In this case, DLLs with direct or indirect circular dependancies will cause you an insane amount of headaches. The best solution is to break the circular dependancies.
P.S.: Note that in C++, constructor can throw, and you don't want an exception in the middle of a DLL loading, so be sure your global objects won't be using exception without a very, very good reason. As correctly written destructors are not authorized to throw, the DLL unloading should be ok in this case.
This page from Microsoft goes into the details of DLL initialization and destruction of globals:
http://msdn.microsoft.com/en-us/library/988ye33t.aspx
If you want to see the actual code that gets executed when linking a .dll, take a look at %ProgramFiles%\Visual Studio 8\vc\crt\src\dllcrt0.c.
From inspection, destructors will be called via _cexit() when the internal reference count maintained by the dll CRT hits zero.
It should be called when either the application ends or the DLL is unloaded, whichever comes first. Note that this is somewhat dependent on the actual runtime you're compiling against.
Also, beware non-trivial destructors as there are both timing and ordering issues. Your DLL may be unloaded after a DLL your destructor relies on, which would obviously cause issues.
In windows binary image files with extension *.exe, *.dll are in PE format
Such files have Entry Point. You can view it with dumpbin tool like
dumpbin /headers dllname.dll
If you use C runtime from Microsoft, then your entry point will be something like
*CRTStartup or *DllMainCRTStartup
Such functions perform initialization of c and c++ runtime and delegate execution to (main, WinMain) or to DllMain respectively.
If you use Microsofts VC compiler then you can watch at source code of this functions in yours VC directory:
crt0.c
dllcrt0.c
DllMainCRTStartup process all things need to init/deinit your global variables from .data sections in normal scenario, when it retrive notification DLL_PROCESS_DETACH during dll unload. For example:
main or WinMain of startup thread of program returns control flow
you explictly call FreeLibrary and use-dll-counter is zero
When DllMain with fdwReason = DLL_PROCESS_DETACH parameter is called it means the DLL is unloaded by the application. This is the time before the destructor of global/static objects gets called.